Data Centers Lose Their Safe Harbor

Ukrainian drones hit Yandex data centers in the Sasovo and Kaluga regions on October 8 and October 9, Ars Technica reported, citing Reuters. Two of the company's five data centers are now knocked out. The AI industry has spent three years asking whether it can find enough chips and power. This week showed a harder question: whether it will be allowed to operate the buildings, and whether anyone will fund them.

That is the argument here. The binding constraint on the buildout is shifting from supply to consent, and consent now comes in three forms: physical security, local permission and capital-market confidence. All three tightened in the same week.

Buildings as targets

Yandex, which Ars Technica says is known as "Russia's Google," is now struggling to keep its AI chatbot and other services running. The Insider reports that Sasovo houses two of Yandex's three supercomputers, which train models including YandexGPT. The company's cloud status page showed the zone tied to Sasovo as completely unavailable and told customers to shift workloads elsewhere. Data Center Dynamics reports that Yandex says it cannot confirm whether the Sasovo equipment can be restored.

Kaluga is the second wound. The Insider found that Yandex had called it its most powerful data center in Russia, and that the original project was designed for more than 3,800 server racks and 63 MW. Meduza reports that after the Sasovo strike, users saw disruptions at Avito, Ivi, Megafon, Russian Railways and T-Bank. One hit propagated across the economy.

The strikes were a reply. Ars Technica says they answered a Russian campaign against Ukraine's data centers that has disrupted mobile and internet service for thousands of Ukrainians since late September. Reuters, via The Moscow Times, reports outages for about 100,000 households in Kyiv and its surroundings. Meduza counts at least eight Ukrainian companies with damaged or destroyed facilities, and says Russia's Defense Ministry has called its strikes a "forced digital detox."

Bloomberg, per MIT Technology Review's The Download, now describes data centers as big targets for military strikes. Nobody's AI roadmap assumed that.

Permission closes in at home

Peacetime operators face a softer version of the same message. San Francisco's Board of Supervisors this week approved a 45-day moratorium on new data centers, the San Francisco Examiner reports. Planning officials must report back within 25 days on defining a data center and studying energy, water, air quality, land use and community effects. The pause could be extended for up to two years.

Supervisor Shamann Walton, who proposed it, called it a "first step toward permanent restrictions on the facilities." Oakland's City Council adopted its own 45-day pause on Tuesday. Walton says a growing list of cities, counties and states have paused or banned new development this year. A moratorium rarely stops a project on its own. It adds delay and uncertainty to every financing model, which is the point.

Capital gets choosy

Then there is the money. Firmus, backed by Nvidia, pulled its ASX listing, ABC News reports, scrapping what would have been the biggest float on the Australian share market since Telstra in 1997. Its co-chief executives targeted a $44 billion valuation. Investors balked at the $11 share price and at bubble worries.

The balance sheet explains why. ABC reports Firmus expects to carry about $43 billion of debt once its data centres are built, set against the operating earnings it forecasts for 2028. Only two of its seven AI factories are operating. UniSuper's chief investment officer, John Pearce, put it bluntly: "It just doesn't have a compelling valuation." Morningstar's Lochlan Holloway described the neo-cloud model as borrowing against customer contracts to buy chips, then using the rent to repay the loans. That works until a customer wobbles. The Guardian, relayed by MIT Technology Review, reports that OpenAI has cut its revenue forecast by $20 billion.

Brookings adds the structural worry. Its new paper projects AI investment in data centers, power and chips could total $10.3 trillion from 2025 to 2032, or 3.63% of U.S. GDP per year. Financing is moving off corporate balance sheets into joint ventures, private credit and special-purpose vehicles, which makes the risk harder to see. Economist Jón Steinsson notes on the Brookings podcast that Meta and Microsoft now spend more on AI than they bring in through operating cash flow.

The main counter-argument

Demand is not the problem, and the main rebuttal starts there. Semiconductor Engineering, citing SIA/WSTS, reports global chip sales hit $159.7B in August, up 8% from July, with year-to-date sales topping $1T. McKinsey expects inference to make up about 60% of AI compute by 2030. Brookings itself says it is premature to call the buildout a systemic risk. Money is still flowing, and a Sydney float is one deal.

The industry also has a workaround. Google says Gemini was trained synchronously across clusters in multiple locations, and Microsoft has linked AI data centres in Wisconsin and Georgia into what it calls one distributed supercomputer. A Cisco-sponsored feature in The Register says spreading compute lets operators build where power is available and planning constraints are lighter.

But the workaround concedes the thesis. It exists because sites are hard to get, and it brings its own bill: Cisco estimates linking two 100MW sites could need 12,000 to 32,000 coherent optical ports, against roughly 1,000 to 2,000 for conventional links. Cisco's Itamar Gold says of multi-site training, "We are seeing early deployments now, but it is a process and I think it will take a few years." Dispersal also trades one risk for another. It creates more sites to permit, more assets to defend and more debt to explain to the people who pulled out of Firmus.

What follows

If this reading holds, the winners will be those who treat location as a core asset, not a procurement line. That means hardened and redundant sites, jurisdictions that say yes in writing, and capital structures that survive an investor asking where the debt sits. Disclosure matters more than ever; Brookings says the time to improve it is while the capital structure is still forming.

Yandex's status page gave customers a simple instruction: move to another availability zone. The rest of the industry is learning how few of those there are, and that chips are the easy part to find.